Fault detection for industrial processes based on dual‐dynamic‐model parallel architecture

Abstract The dynamic characteristics of industrial processes and complex coupling relationships among variables greatly limit fault detection performance. To address this issue, a dual‐dynamic‐mode parallel architecture based on causal partitioning is proposed in this paper. First, the causal driving relationships between input and output variables are analyzed using transfer entropy, thereby enabling the intelligent and accurate partition of input variables into output‐related and output‐unrelated sets. Second, the dynamic principal component analysis (DPCA) model is established for output‐unrelated variables. Then, by decoupling the dynamic model between output‐related and output variables, it is separated into external‐driving and self‐driving components. Finally, the DPCA model, external‐driving and self‐driving sub‐models are then integrated, collectively forming a dual‐dynamic‐mode parallel architecture, which enables comprehensive and accurate fault detection and assists in fault causal identification. Using a numerical example and an actual thermal power plant case, the proposed method is rigorously tested, which effectively verifies its strong applicability and distinct advantages.